· 07:00 · 7 min + 2 min Explained
Mistral's Voxtral 3 turns speech into text in 50 languages
A European lab now sells cheap transcription in every official EU language. Also: TorchTPU and DeepInfer loosen NVIDIA's grip, Anthropic adds user profiles, ChatGPT gets chat locks, and TypeScript in Rust.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- Mistral's Voxtral 3 transcribes 50 languages, including all 24 official EU languages, from $0.0035 a minute. Anyone who turns meetings, calls or interviews into text.
- Google open-sources TorchTPU, and DeepSeek shares the software that runs its own models. Teams that want to run AI without depending on NVIDIA.
- Anthropic lets platforms tag each end customer, and pitches a law to halt dangerous models. Apps that resell Claude, and anyone following AI rules.
- ChatGPT can lock chats behind a PIN or Face ID, and OpenAI buys a drug-trial agent startup. Anyone who shares a phone or laptop, and pharma teams.
- Cloudflare and partners port the TypeScript compiler to Rust, up to 8 times faster. Web developers who wait on slow type checks.
1. Mistral’s Voxtral 3 turns speech into text in 50 languages
The short versionFrench lab Mistral released Voxtral 3, a family of speech models. It transcribes all 24 official EU languages, matches the best rivals on Mistral's own test, and costs fractions of a cent per minute.
Speech to text turns meetings into notes and interviews into quotes. But every wrong word means someone has to fix the text by hand, so the error rate decides how useful it is.
Mistral has launched Voxtral 3. Its main transcription models now cover 50 languages, including all 24 official EU languages.
On a standard test in 45 languages, its larger model gets about 4 words wrong in 100. OpenAI’s Whisper large-v3 gets 16 wrong. The small model costs $0.0035 per minute.
Mistral offers it through a European endpoint as well as US and global ones.
There is a catch. These are Mistral’s own charts, and they show Microsoft’s model level with it and others close behind. The page also does not say whether anyone can download the models.
Whisper is like a typist who needs a lot of correcting. Voxtral is a better typist for a small fee. For a Danish clinic or law firm, that means far less fixing, and the audio can stay in Europe.
Try it: Mistral's API has Voxtral 3 today. Test it on a recording in your own language before you switch.
Source: mistral.ai · Official
2. Two open releases make it easier to run AI without NVIDIA
The short versionGoogle made TorchTPU open source, so PyTorch code can run on its chips. DeepSeek shared the software that serves its own models, and it runs on Huawei chips too.
Most AI software is written for NVIDIA’s chips, so switching to other chips takes real work. Two new releases make it easier.
1 of 2 Google opens up TorchTPU
PyTorch is the toolkit most AI models are built with. Google built TorchTPU with Meta’s PyTorch team, so PyTorch code can run on Google’s own TPU chips with “minimal code changes”. In version 5 of Hugging Face’s Transformers library, you pick the TPU with a single setting.
2 of 2 DeepSeek shares its serving engine
DeepSeek released DeepInfer, the software that runs its models in production, under the free Apache 2.0 license. It works on NVIDIA’s Blackwell (SM100) chips and on Huawei’s Ascend 950. For now it only runs one model, DeepSeek-V4.1-Flash, and some production parts are left out, such as a memory cache.
Chips are like electrical sockets: a plug that fits only one socket ties you to one supplier. Both releases are adapters. More adapters make it harder for one chipmaker to set the price.
Source: cloud.google.com · Official · github.com · Official
3. Anthropic adds user profiles to its API and pitches a safety law
The short versionApps that resell Claude can now tell Anthropic which customer each request is for. Separately, Anthropic is showing Congress a draft law to stop a dangerous model's release.
1 of 2 One bad customer no longer has to sink a whole app
Many apps run on Claude behind the scenes. If one of their users broke Anthropic’s rules, enforcement could hit the whole app. With user profiles, launched on 20 October, a platform tags each request with its end customer. Then enforcement can hit that one customer instead.
2 of 2 A draft law with a court as the brake
Anthropic has shown lawmakers a draft bill, TechCrunch reports. Independent testers would check new models. If they find an imminent risk of over 50 deaths or $1 billion in damage, they have 72 hours to tell the Commerce Department. Only a judge could then halt the release. Labs could pay the testers, and TechCrunch asks how that would not sway results. Anthropic confirmed the draft to Politico.
User profiles are like a landlord who can evict one noisy tenant instead of closing the building. The bill tries the same thing for whole models: a narrow brake, used by a court, only for a disaster.
Do this: If you resell Claude to your own customers, add user profiles to your API calls.
Source: platform.claude.com · Official · techcrunch.com · Reported
4. ChatGPT can lock your chats, and OpenAI buys a drug-trial agent startup
The short versionChatGPT now hides chosen chats behind a PIN or Face ID. OpenAI also bought Orchestra Bio to add AI agents to its tools for drug companies.
1 of 2 Locked chats stay private on a shared screen
Many people ask ChatGPT about health or money, then hand their phone to someone. OpenAI began rolling out a lock on 20 October. You can lock a chat with a four-digit PIN or Face ID, TechCrunch reports. It works on most plans, from Free to Enterprise.
The lock is a curtain, not a safe. It adds no extra encryption, and it does not change what OpenAI keeps or trains on. If you reset the PIN, all locked chats are deleted for good.
2 of 2 OpenAI moves deeper into drug development
OpenAI bought Orchestra Bio, which builds agents that track budgets and fix scheduling problems in drug trials. The team joins Rosalind, OpenAI’s AI model for life science research. OpenAI says Moderna and Denmark’s Novo Nordisk use Rosalind.
The lock hides a chat from someone holding your phone, but OpenAI can still read it. The most sensitive things still do not belong in a chat.
Source: techcrunch.com · Reported · techcrunch.com · Reported
5. The TypeScript compiler gets a Rust port that runs up to 8 times faster
The short versionCloudflare and partners released tsrs, an experimental TypeScript compiler written in Rust. They say it is up to 8 times faster than TypeScript 7.0.
In big web projects, checking for type errors can be slow. On 20 October, Cloudflare, VoidZero and Owner.com’s Max Schwenk released tsrs. It is built with the TypeScript team, which keeps control of the language.
The team says tsrs is up to 8 times faster than TypeScript 7.0 and uses about 70% less memory. TypeScript’s own 7.1 beta is already up to 2.5 times faster than 7.0, so the real gap is smaller. Cloudflare also writes that frontier AI models can now do ports like this almost on their own in days. It is still experimental, and some output differs from the original.
If AI can do a port like this in days, the tools developers wait on every day can get faster much sooner.
Try it: Run npx tsrs on a project and compare the time with your usual check.
Source: blog.cloudflare.com · Official
Open source & local AI
PyTorch 2.15 runs mixture-of-experts models up to 6.5 times faster on Macs
The short versionThe next PyTorch makes Apple Silicon Macs much faster at running many open models locally. It is a test release for now.
Many new open models are mixtures of experts: they only use a small part of themselves for each word. On a Mac, PyTorch handled them badly, so they ran slowly.
Hugging Face’s PyTorch team says version 2.15 fixes that. Such models now write their answers 5 to 6.5 times faster than in 2.14. In the best case, on an M3 Ultra, Qwen3.6-35B-A3B went from 9 to 61 tokens per second. Other settings in the post show smaller gains.
It is a test version for now, and it needs macOS 15 or newer.
Nine tokens per second is like waiting for someone to type. Sixty is faster than you can read. That gap decides whether a private model on your own Mac feels usable at all.
Source: huggingface.co · Official
Paper of the day
Shrinking an agent’s memory is often tested on the wrong question
The short versionWays of trimming a long conversation look good on the question at hand. On the questions that come later, many do no better than deleting sentences at random.
AI agents run out of room in long sessions. So they compact: they shrink the history and keep what seems important. But what seems important now may not be what you ask about next.
A team led by Junwei Yao tested this. Keeping only the sentences needed for the current question raised accuracy on that question by 11.4 points, at a quarter of the text.
On later questions, no such method did measurably better than random deletion. LLMLingua-2, a tool that ignores the question, beat random deletion on later questions in one test.
On agent tasks, where later tool calls needed exact IDs, choosing sentences for the current question did beat random deletion.
It is like packing for a trip by only thinking about the first day. The lesson: test compaction on what comes next, not on what you just asked.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 2 minWhat is word error rate?
Word error rate, or WER, is a score for how well a speech-to-text model writes down what people say. It counts the words the model gets wrong. To measure it, you need a recording and a correct transcript written by a person. The model's text is compared with that transcript, word by word. Three kinds of mistakes count. A wrong word is a substitution. A missing word is a deletion. An extra word is an insertion. You add up all the mistakes and divide by the number of words in the correct transcript. > **Think of it like this:** a teacher marks a dictation. Every wrong, missing or extra word gets a red mark. WER is the number of red marks divided by the number of words in the text. Here is an example from Hugging Face's course. The correct text is "the cat sat on the mat", six words. The model writes "the cat sit on the". That is one wrong word and one missing word, so two mistakes out of six, or about 33%. Lower is better, and a perfect model scores 0%. WER can even go above 100%, if a model adds many extra words. Scores also depend on the test. Removing capital letters and punctuation first makes WER lower. So two scores are only comparable when they come from the same test with the same rules. **Why it matters to you:** when a company says its model has a lower WER, it means fewer words you have to fix by hand. Check that the test includes your language, because scores can be very different from one language to the next.Two more from today
What is a mixture-of-experts (MoE) model?Why huge models only use a slice of themselves for each word.
A mixture-of-experts model is split into many smaller specialist parts, called experts. For each word it writes, a small router picks only a few of them to do the work.
So a model can have hundreds of billions of weights in total but use only a fraction per word. That makes it much faster and cheaper to run than a dense model of the same size.
What is a context window?How much a model can read at once, and why it runs out.
The context window is how much text a model can look at in one go: your question, the documents you paste in and its own answer. It is measured in tokens, roughly three quarters of a word each.
Anything outside the window the model simply cannot see. That is why long chats get summarised before they run out of room.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak